Introduction to Machine Learning
A fast conceptual map of the field and why learning from data works.
What machine learning is
Machine learning (ML) is the study of algorithms that improve their behavior on a task using data. Instead of hand-coding every rule, we define a model family and let optimization find parameter values that perform well on examples.
The standard framing is: a task T, performance measure P, and experience E. A system learns if performance at T, measured by P, improves with E.
Core learning settings
- Supervised learning: learn from labeled pairs \\(x, y\\) for classification or regression.
- Unsupervised learning: discover structure (clustering, representation learning, density estimation).
- Self-supervised learning: generate targets from data itself and learn useful features.
- Reinforcement learning: optimize long-term reward through interaction.
Generalization, not memorization
Success in ML is measured on unseen samples. This is why train/validation/test splits exist: we need evidence that the learned rule captures regularities of the data-generating process rather than accidental quirks of the training set.
Bias-variance trade-offs, regularization, and model capacity all control this generalization behavior.
The modern training pipeline
- Define objective and metric.
- Collect and clean data.
- Choose model class and loss.
- Optimize parameters (usually gradient-based).
- Evaluate, diagnose errors, and iterate.
Takeaway: ML is a loop between modeling assumptions, optimization, and data quality. Understanding all three is more important than chasing any single algorithm.